How to meet asynchronously (almost) everywhere Jurek Czyzowicz Arnaud Labourel Andrzej Pelc

Transcription

How to meet asynchronously (almost) everywhere Jurek Czyzowicz Arnaud Labourel Andrzej Pelc
How to meet asynchronously (almost) everywhere
Jurek Czyzowicz
∗†
Arnaud Labourel
∗‡
Andrzej Pelc
∗§
Abstract
Two mobile agents (robots) with distinct labels have to meet in an arbitrary, possibly infinite, unknown connected graph or in an unknown connected terrain in the plane. Agents are
modeled as points, and the route of each of them only depends on its label and on the unknown
environment. The actual walk of each agent also depends on an asynchronous adversary that
may arbitrarily vary the speed of the agent, stop it, or even move it back and forth, as long as
the walk of the agent in each segment of its route is continuous, does not leave it and covers all
of it. Meeting in a graph means that both agents must be at the same time in some node or in
some point inside an edge of the graph, while meeting in a terrain means that both agents must
be at the same time in some point of the terrain. Does there exist a deterministic algorithm
that allows any two agents to meet in any unknown environment in spite of this very powerfull
adversary? We give deterministic rendezvous algorithms for agents starting at arbitrary nodes of
any anonymous connected graph (finite or infinite) and for agents starting at any interior points
with rational coordinates in any closed region of the plane with path-connected interior. While
our algorithms work in a very general setting – agents can, indeed, meet almost everywhere –
we show that none of the above few limitations imposed on the environment can be removed.
On the other hand, our algorithm also guarantees the following approximate rendezvous for
agents starting at arbitrary interior points of a terrain as above: agents will eventually get at
an arbitrarily small positive distance from each other.
keywords: rendezvous, graph, asynchronous, deterministic
∗
D´epartement d’informatique, Universit´e du Qu´ebec en Outaouais, Gatineau, Qu´ebec J8X 3X7, Canada. E-mails:
[email protected], [email protected], [email protected]
†
Partially supported by NSERC discovery grant.
‡
This work was done during this author’s stay at the Universit´e du Qu´ebec en Outaouais as a postdoctoral fellow.
§
Partially supported by NSERC discovery grant and by the Research Chair in Distributed Computing at the
Universit´e du Qu´ebec en Outaouais.
1
Introduction
The problem and the model. Two mobile agents (robots) modeled as points starting at different
locations of an unknown environment have to meet. This task is known in the literature as the
rendezvous problem, and has been studied under two alternative scenarios. Either the agents move
in a network, modeled by an undirected connected graph (the graph scenario), or they move in
(some subset of) the plane (the geometric scenario).
In this paper we study the asynchronous version of the rendezvous problem, under both above
scenarios: each agent designs its route and an adversary controls the speed of each agent, can vary
this speed, stop the agent, or even move it back and forth, as long as the walk of the agent in each
segment of its route is continuous, does not leave it and covers all of it. In the asynchronous version
of the graph scenario, meeting at a node may be impossible even in the two-node graph, as the
adversary can desynchronize the agents and make them visit nodes at different times. Thus it is
necessary to relax the requirement and allow agents to meet either in a node or inside an edge. Such
a definition of meeting is natural, e.g., when agents are robots traveling in a labyrinth. We consider
an embedding of the underlying graph in the three-dimensional Euclidean space, with nodes of the
graph being points of the space and edges being pairwise disjoint line segments joining them (hence
there are no edge crossings). Agents are modeled as points moving inside this embedding.
If nodes of the graph are labeled and the labeling is known, then agents can decide to meet at a
predetermined node and the rendezvous problem reduces to graph exploration. However, in many
applications, when rendezvous is needed in a network of unknown topology, such unique labeling
of nodes may be unavailable, or agents may be unable to perceive such labels, e.g., due to security
reasons. Hence it is important to design rendezvous algorithms for agents operating in anonymous
graphs, i.e., graphs without unique labeling of nodes. It is important to note that the agents have
to be able to locally distinguish ports at a node: otherwise, an agent may even be unable to visit
all neighbors of a node of degree 3 (after visiting the second neighbor, the agent cannot distinguish
the port leading to the first visited neighbor from that leading to the unvisited one). Consequently,
agents initially located at two nodes of degree 3, might never be able to meet. This justifies a
common assumption made in the literature: all ports at a node are locally labeled by distinct
positive integers. Degrees of nodes can be either finite or infinite. No coherence between those
local labelings is assumed. When an agent leaves a node, it is aware of the port number by which
it leaves and when it enters a node, it is aware of the entry port number. It can also verify, at
each node, whether a given positive integer is a port number at this node. Agents know neither the
graph, nor the initial distance between them. They cannot mark the nodes or the edges in any way.
Rendezvous has to be accomplished regardless of local labelings of ports. Each agent terminates
its walk at the time of meeting the other agent.
In the geometric scenario, we assume that the terrain in which agents operate is a closed subset of
the Euclidean plane, i.e., it contains limits of all converging sequences of points in it. The boundary
of the terrain is defined as the set of points having arbitrarily close points both in the terrain and
outside of it. Since the terrain is closed, the boundary is included in it. All other points of the
terrain are its interior points. Each agent can only distinguish if it is currently in an interior point
of the terrain or in its boundary. Agents do not know the terrain in which they operate, they
cannot “see” any vicinity of the currently visited point and they cannot leave any marks. Agents
are equipped with a compass and with the same unit of length. Thus systems of coordinates of
1
agents are aligned and the origin for each agent is at its starting point. Again, an agent terminates
its walk at the time of meeting the other agent.
If agents are identical, i.e., they do not have distinct identifiers, and execute the same algorithm,
then deterministic rendezvous is impossible, e.g., in the ring (graph scenario) or in the plane
(geometric scenario): the adversary will make the agents move always in the same direction at the
same speed, keeping them at the same distance at all times, thus they will never meet. Hence we
assume that agents have distinct identifiers, called labels, which are two different positive integers.
This is their only way to break symmetry. We assume that each agent knows its own label but
not the label of the other agent. This excludes, e.g., rendezvous strategies of the type “waiting
for mommy”, in which the agent with smaller label remains idle and the other agent explores the
graph or the terrain. We do not impose any restriction on the memory of the agents: from the
computational perspective they are viewed as Turing machines.
Two important notions used to describe movements of agents are the route of each agent and its
walk. Roughly speaking, each agent chooses the route where it moves and the adversary describes
the walk on this route, deciding how the agent moves. More precisely, these notions are defined as
follows. The adversary initially places an agent with label ` at some node of the graph or at some
point in the terrain. Given this label and this starting point, the route is chosen by the agent and
is defined as follows. In the case of the graph, the agent chooses one of the available ports at the
current node. After getting to the other end of the corresponding edge, the agent chooses one of
the available ports at this node, and so on, indefinitely (until rendezvous). The resulting route of
the agent is the corresponding sequence of edges, which is a (not necessarily simple) path in the
graph. The route in a terrain is a sequence (S1 , S2 , . . . ) of segments, where Si = [ai , ai+1 ], defined
in stages as follows, given the agent’s label and the starting point. In stage i the agent starts at
point ai , and a1 is the starting point chosen by the adversary. The agent chooses a direction α and
distance x. If the segment of length x in direction α starting in v intersects the boundary of the
terrain at some distance y ≤ x from v, the agent becomes aware of it in the intersection point w
closest to v. In this case, the stage ends at w and in the next stage the agent chooses the reverse
direction α and the distance y (that will cause it to return to v). If the segment of length x in
direction α starting in v does not intersect the boundary of the terrain, the stage ends when the
agent reaches point u at distance x from v in direction α. Stages are repeated indefinitely (until
rendezvous).
We now describe the walk f of an agent on its route. Let R = (S1 , S2 , . . . ) be the route of an agent.
In the graph scenario this is a (not necessarily simple) infinite path in the (spatial embedding of)
the graph, and in the geometric scenario it is an infinite polygonal line in the plane. Let (t1 , t2 , . . . ),
where t1 = 0, be an increasing sequence of reals, chosen by the adversary, that represent points in
time. Let fi : [ti , ti+1 ] → [ai , ai+1 ] be any continuous function, chosen by the adversary, such that
fi (ti ) = ai and fi (ti+1 ) = ai+1 . For any t ∈ [ti , ti+1 ], we define f (t) = fi (t). The interpretation
of the walk f is as follows: at time t the agent is at the point f (t) of its route. This general
definition of the walk and the fact that it is constructed by the adversary capture the asynchronous
characteristics of the process. The movement of the agent can be at arbitrary speed, the agent may
sometimes stop or go back and forth, as long as the walk in each segment of the route is continuous
and covers all of it.
Notice that the power of the asynchronous adversary to produce any continuous walk on the routes
determined by the agents implies the following significant difference with respect to the synchronous
scenario. While in the latter scenario the relative movement of the agents depends only on their
routes, in our setting, this movement is also controlled by the adversary.
2
Agents with routes R1 and R2 and with walks f1 and f2 meet at time t, if points f1 (t) and f2 (t) are
identical. A rendezvous is guaranteed for routes R1 and R2 , if the agents using these routes meet
at some time t, regardless of the walks chosen by the adversary. A rendezvous algorithm executed
by agents in a graph or in a terrain produces routes of agents, given the label of each agent and its
starting point.
It should be stressed that, while routes of agents are formally defined as infinite sequences of
segments, our results imply that in any instance of the rendezvous problem, meeting will occur
at some finite time, and thus each agent will compute only finitely many segments of its route.
As mentioned above, agents compute their routes in stages, and given any walk chosen by the
adversary, each stage is completed in finite time. There is no stopping issue in our solution:
rendezvous always occurs at some stage for each of the agents and then both agents stop. Another
feature of our rendezvous algorithms is that in the choice of consecutive segments of its route an
agent does not use the knowledge of the walk to date. Thus the route depends only on the label of
the agent, on the environment (graph or terrain), and on the starting point chosen by the adversary,
but not on its other actions.
Our results. We give two deterministic algorithms, the first for rendezvous in the graph scenario
and the second in the geometric scenario. For the graph scenario, our algorithm accomplishes
rendezvous in any connected countable∗ (finite or infinite) graph, for arbitrary starting nodes. A
consequence of this very general result is the positive answer to the following question from [12]: Is
deterministic asynchronous rendezvous feasible in any finite connected graph without knowing any
upper bound on its size? (In [12] the authors presented a deterministic asynchronous rendezvous
algorithm in arbitrary finite connected graphs with known upper bound on the size.)
For the geometric scenario, our algorithm accomplishes rendezvous for agents starting at any interior
points with rational coordinates in any closed region of the plane with path-connected interior.
(Recall that a subset T of the plane is path-connected, if for any points u, v ∈ T , there is a
continuous function h : [0, 1] −→ P , such that P ⊆ T and h(0) = u, h(1) = v.) On the other hand,
our algorithm guarantees the following approximate rendezvous for agents starting at arbitrary
interior points of a terrain as above: agents will eventually get at an arbitrarily small positive
distance from each other. This implies the perhaps surprising result that if agents have arbitrarily
small positive visibility ranges (rather than 0 visibility range as we assume) and they start in
arbitrary points of the (empty) plane, then they will see each other in finite time, regardless of the
actions of the adversary.
Discussion of limitations. While our algorithms work in a very general setting – agents can,
indeed, meet almost everywhere – it turns out that none of the few limitations imposed on the
environment can be removed. For the graph scenario, the only limitation is connectivity of the
graph. It is clear that rendezvous in disconnected graphs is impossible, if the agents start in
different connected components. For the geometric scenario, let us review the limitations one by
one. First, we assume that the terrain is closed. This assumption cannot be entirely removed for
the following technical reason. Consider the construction of a route in an open disc. An agent
starting at any point, that chooses in the first stage an arbitrary direction and a sufficiently large
distance, at some point would have to leave the disc. Since it does not see anything in its vicinity,
it cannot know where the boundary is before hitting it, and it cannot hit it, as it is not allowed to
leave the terrain. It follows that the agent could not construct further segments of its route. The
second assumption is that agents start at interior points of the terrain. This assumption cannot
∗
A graph is countable if the set of its nodes is countable, i.e., if there exists a one-to-one function from this set
into the set of natural numbers.
3
be removed either. Indeed, suppose that the terrain is a closed disc with a semi-circle attached
to it. This is a closed subset of the plane with nonempty path-connected interior. Suppose that
one agent starts in the disc and the other at the end of the semi-circle. Since agents need to move
along polygonal lines, the second agent could not move at all and the first one cannot reach it. Our
next assumption is that the interior of the terrain is path-connected. To show that this assumption
cannot be removed, consider two disjoint closed discs joined by an arc of a circle. This terrain is
closed and path-connected, but if each agent starts inside a different disc, again they cannot meet,
because agents need to move along polygonal lines, and hence cannot traverse the joining arc. The
final assumption is that the starting points of the agents have rational coordinates. In Section 4 we
prove that if the agents start in arbitrary points, then rendezvous cannot be guaranteed even in the
plane. We show, however, that for arbitrary starting points approximate rendezvous is guaranteed.
Related work. The rendezvous problem was first described in [25]. A detailed discussion of the
large literature on rendezvous can be found in the excellent book [4]. Most of the results in this
domain can be divided into two classes: those considering the geometric scenario (rendezvous in
the line, see, e.g., [9, 10, 17], or in the plane, see, e.g., [7, 8]), and those discussing rendezvous in
graphs, e.g., [2, 5]. Some of the authors, e.g., [2, 3, 6, 9, 18] consider the probabilistic scenario
where inputs and/or rendezvous strategies are random. Randomized rendezvous strategies use
random walks in graphs, which were thoroughly investigated and applied also to other problems,
such as graph traversing [1], on-line algorithms [13] and estimating volumes of convex bodies [15].
A generalization of the rendezvous problem is that of gathering [16, 18, 19, 20, 23, 27], when more
than 2 agents have to meet in one location.
If graphs are unlabeled, deterministic rendezvous requires breaking symmetry, which can be accomplished either by allowing marking nodes or by labeling the agents. Deterministic rendezvous
with anonymous agents working in unlabeled graphs but equipped with tokens used to mark nodes
was considered e.g., in [22]. In [28] the authors studied gathering many agents with unique labels.
In [14, 21, 29] deterministic rendezvous in graphs with labeled agents was considered. However, in
all the above papers, the synchronous setting was assumed. Asynchronous gathering under geometric scenarios has been studied, e.g., in [11, 16, 24] in different models than ours: agents could
not remember past events, but they were assumed to have at least partial visibility of the scene.
The first paper to consider deterministic asynchronous rendezvous in graphs was [12]. The authors
concentrated on complexity of rendezvous in simple graphs, such as the ring and the infinite line.
They also showed feasibility of deterministic asynchronous rendezvous in arbitrary finite connected
graphs with known upper bound on the size. Further improvements of the above results for the infinite line were proposed in [26]. Gathering many robots in a graph, under a different asynchronous
model and assuming that the whole graph is seen by each robot, has been studied in [19, 20].
2
Preliminary notions and results
A fundamental notion on which our algorithms are based is that of a tunnel. Consider any graph G
and two routes R1 and R2 starting at nodes v and w, respectively. We say that these routes form a
tunnel, if there exists a prefix [e1 , e2 , . . . , en ] of route R1 and a prefix [en , en−1 , . . . , e1 ] of route R2 ,
for some edges ei in the graph, such that ei = {vi , vi+1 }, where v1 = v and vn+1 = w. Intuitively,
the route R1 has a prefix P ending at w and the route R2 has a prefix which is the reverse of
P , ending at v. By a slight abuse of terminology we will also say that prefixes [e1 , e2 , . . . , en ] and
[en , en−1 , . . . , e1 ] form a tunnel.
Proposition 2.1 If routes R1 and R2 form a tunnel, then they guarantee rendezvous.
4
Proof: Consider an embedding of the graph G in the three-dimensional Euclidean space, with
nodes of the graph being points of the space and edges being pairwise disjoint line segments joining
them. Consider routes R1 and R2 starting at nodes v and w, respectively. Let agent ai execute
route Ri . Let P be the polygonal line joining v with w, corresponding to the prefixes of the routes,
given by the tunnel. Let D be its length defined as the sum of lengths of edges in the corresponding
prefixes of the routes. (For non-simple paths in the graph, the same edge is counted many times.)
Consider any walks f1 on R1 and f2 on R2 . Let t0 be the first moment when an agent leaves its
starting point and let t00 be the moment when an agent gets to the end of P other than its starting
point. For any t ∈ [t0 , t00 ], let d1 (t) be the distance of agent a1 from its starting point v at time
t, counted on the route R1 , and let d2 (t) be the distance of agent a2 from its target point v at
time t, counted on the route R2 . Let δ(t) = d2 (t) − d1 (t). We have δ(t0 ) = D and δ(t00 ) = d ≤ 0.
The function δ is thus a continuous function from the interval [t0 , t00 ] onto some interval [d0 , D],
where d0 ≤ d, in view of the continuity of walks f1 and f2 . Since 0 belongs to the interval [d0 , D],
there must exist a moment t in the interval [t0 , t00 ], for which δ(t) = 0. For this point we have
f1 (t) = f2 (t), and the rendezvous occurs.
We now recall some basic facts from set theory, that will be used in further considerations.
Proposition 2.2 The set of rational numbers and the set of positive rational numbers are countable. The cartesian product of two countable sets is countable. The set of all finite sequences with
terms in a countable set is countable.
3
Rendezvous in the graph scenario
Let G = (V, E) be the connected graph in which the rendezvous must be performed. Let Sn be the
set of sequences of n positive integers. Let P = {(i, j, s0 , s00 ) | i, j ∈ N, i < j and ∃n s.t. s0 , s00 ∈ Sn }.
There exists a bijection from the set of positive integers onto P, in view of Proposition 2.2. Let
(ϕ1 , ϕ2 , . . .) be a fixed enumeration of P. All agents have to agree on the same enumeration. It
is easy to produce a formula computing φk for any k. This formula is included in the rendezvous
algorithm. For a finite path r in G, we denote by r the path with the same edges as in r, but in
the reverse order. Remark that r and r form a tunnel.
We first give a high-level idea of the algorithm referring to lines of the pseudo code given below.
We “force” the routes of any two agents to form a tunnel for every possible combination of starting
nodes and labels of the two agents. By Proposition 2.1, this suffices to guarantee rendezvous. Any
starting configuration of robot i placed at node v and robot j placed at node w by the adversary
corresponds to a quadruple (i, j, s0 , s00 ) where s0 is a sequence of ports inducing a path from v to w
and s00 is a sequence of ports inducing the reverse path from w to v.
Each agent constructs its route in phases. At the beginning and at the end of each phase the
agent is in its starting node. At phase k the previously constructed initial part of the route rhist is
extended while the agent processes quadruple ϕk (some of the extensions are null). This extension
guarantees that the routes of agents of the corresponding starting configuration will form a tunnel.
When agent with label l processes quadruple ϕk = (i, j, s0 , s00 ) nothing happens if l 6= i and l 6= j
(line 4). If l = i, agent i tries to extend its route to guarantee rendezvous with agent j under
the hypothesis that a path from v to w corresponds to the sequence s0 of ports and the reverse
path corresponds to the sequence s00 . For this to happen, the agent first tries to follow the path
r(s0 ) induced by the sequence s0 of ports (lines 8-11). This attempt is considered successful if the
following conditions are satisfied:
5
• at consecutive nodes of the traversed path, ports with numbers from the sequence s0 are available,
• the reverse path corresponds to the sequence s00 of ports.
When the attempt is successful (the condition of line 13 is satisfied) the agent is at node w and
it simulates the first k − 1 phases of the execution of the algorithm by agent with label j starting
from w. The effect of this simulation is the path rsim . Upon completion of this part, agent with
label i returns to w. Now the agent is able to further extend its path to form a tunnel with the
route of agent j (line 16). Finally, whether the attempt to follow the path r(s0 ) is successful or not,
the agent with label i backtracks to v (line 17). If l = j, the above actions are performed with the
roles of i and j reversed and the role of s0 and s00 reversed.
Algorithm GraphRV calls the recursive function GraphRVREC. This function is called in two different
modes controlled by the boolean mode. In the “main” mode (mode = true) the function is executed
indefinitely, until rendezvous. In the “simulation” mode (mode = f alse), the function is executed
for all values up to a given p, or until rendezvous, whichever comes first. The symbol a denotes
the concatenation of sequences.
Algorithm GraphRV
INPUT: A starting node v ∈ V and a label l of the agent.
OUTPUT: A rendezvous route r.
GraphRVREC(v, l, 0, true);
function GraphRVREC(node v, label l, integer p, boolean mode)
1 k := 1; r := λ;
2 while not rendezvous and (k ≤ p or mode) do
3
let ϕk = (i, j, s0 , s00 ); rhist := r;
4
if l = i or l = j then
5
if l = i then s1 := s0 ; s2 := s00 ; l0 := j;
6
else s1 := s00 ; s2 := s0 ; l0 := i;
7
let s1 = (p1 , . . . , pn ); m := 1; r(s1 ) := λ;
8
while m ≤ n and pm is a port at the current node do
9
r(s1 ) := r(s1 ) a (em ) where em is the edge corresponding to port pm ;
10
let am be the port corresponding to em at its other endpoint;
11
m := m + 1;
12
r := r a r(s1 );
13
if s2 = (an , . . . , a1 ) then
14
let w be the current node;
15
rsim := GraphRVREC(w, l0 , k − 1, f alse);
16
r := r a rsim a r(s1 ) a rhist a r(s1 ) a rsim ;
17
r := r a r(s1 );
18 k = k + 1;
19 return r
Theorem 3.1 Algorithm GraphRV guarantees asynchronous rendezvous for arbitrary two agents
starting from any nodes of an arbitrary connected graph.
Proof: Let v and i (resp. w and j) be the starting node and the label of the first agent (resp.
the second agent). There exists a path q linking v to w, since the graph G is connected. Let s0
6
rw
w
q
rv
v
route of agent i: rv _ q _ rw _ q _ rv _ q _ rw
route of agent j: rw _ q _ rv _ q _ rw _ q _ rv
tunnel: rv _ q _ rw _ q _ rv _ q _ rw
_
_
q···
q...
Figure 1: Tunnel between the routes of two agents
(resp. s00 ) be the finite sequence of ports corresponding to the path q (resp. the path q). In view of
Proposition 2.1, it suffices to prove that the routes of the two agents form a tunnel. We show that,
after the phase corresponding to the quadruple ϕk = (i, j, s0 , s00 ) during the execution of Algorithm
GraphRV for agents i and j, the routes of the two agents form a tunnel. Observe that this phase
eventually occurs during any execution of GraphRVREC, since all recursive calls of any phase k 0 < k
are done with a parameter p strictly smaller than k 0 . Thus all these phases are completed in finite
time.
First, we show by induction that, at the beginning of phase k of any execution of Algorithm GraphRV,
each agent is at its starting node. This is clearly true for k = 1. Assume that the property holds
for k − 1. It follows that during the execution of the phase k − 1, the paths rhist and rsim are cycles.
Hence, after the execution of line 16, the agent ends in node w. After the execution of line 17, the
agent returns to the starting node v of the phase k − 1. So, the agent starts phase k in the same
node, and the property is true for all k.
Let rv (resp. rw ) be the output of the execution of the first k − 1 phases of Algorithm GraphRV for
agent i (resp. j) starting in node v (resp. w). At the beginning of phase k, the portion of the route
constructed by agent i is rv . After the execution of line 12, the portion of the route constructed by
agent i is rv a q, since the agent has started the phase in node v. The path rsim computed by the
recursive call of GraphRVREC is equal to rw . It follows that at the end of phase k, the portion of the
route constructed by agent i is ρ = rv a q a rw a q a rv a q a rw a q. Similarly, at the end of phase
k, the portion of the route constructed by agent j is ρ0 = rw a q a rv a q a rw a q a rv a q. By
construction, the part rv a q a rw a q a rv a q a rw of ρ and the part rw a q a rv a q a rw a q a rv
of ρ0 form a tunnel (see Fig 1).
4
Rendezvous in the geometric scenario
In this section we consider the problem of rendezvous in a terrain included in the Euclidean plane.
As announced in the introduction, we restrict attention to closed subsets of the plane whose interior
is path-connected. We observed that these restrictions cannot be removed.
Fix a system of coordinates Σ with the y-axis pointing to North (shown by compasses of the agents)
and with the unit of length equal to that of the agents. Points with rational coordinates in Σ will
7
be called rational. A polygonal line all of whose vertices, including extremities, are rational will be
called a rational line. For any point u, let Σu be the shift of the system Σ with origin at point u.
Lemma 4.1 In any path-connected, open subset S of the plane and for any rational points u, v ∈ S,
there exists a rational polygonal line included in S, with extremities u and v, all of whose vertices
are rational.
Proof: By path-connectivity of S, there exists a path p included in S with extremities u and
v, which is a continuous image of the interval [0, 1]. Let d be the distance from p to cS - the
complement of S. Since p ∩ cS = ∅ and both p and cS are closed sets, we have d > 0. Partition the
plane into squares of side length at most d/2 with rational vertices. Let Q be the set of squares
intersecting p. Since p is a bounded set, Q is finite. Consider the graph Gp with node set Q, such
that x, y ∈ Q are adjacent if p contains a point belonging to a common boundary of x and y.
Since Gp is connected, there exists a path (x1 , . . . , xk ) in Gp linking squares x1 containing u and xk
containing v. Let p∗ be the polygonal path (uw1 , w1 w2 , . . . , wk−1 wk , wk v), where wi is the center
of square xi , for all i = 1, . . . , k. The path p∗ is rational
and contained in the union of squares from
√
Q. Since each point of p∗ is at distance at most d 2/2 from some point of p, we have p∗ ∩ cS = ∅,
hence p∗ is included in S.
We define the following graph GT = (V, E), for a given terrain T . The set of nodes V is the union
of two disjoint subsets V1 , V2 . The set V1 is the set of all interior, rational points of T and the set
V2 is defined below.
For each pair of points p1 , p2 , such that p1 ∈ V1 and p2 is any rational point of the plane, we
consider the segment s = p1 p2 . If s does not intersect the boundary of T , then p2 must belong to
V1 and we add the edge {p1 , p2 } to E. If s intersects the boundary of T , we add a new node v to
V2 and we add the edge {p1 , v} to E. Note that such a node v is always added to V2 when point
p2 is on the boundary of T or outside of T (and it may or may not be added to V2 when p2 is in
the interior of T ). Since each node in V2 corresponds to a pair of rational points p1 , p2 , there is a
countable number of nodes in V2 , each of them having degree 1. The unique port at any node in
V2 has number 1 and ports at any node in V1 are defined as follows. Let (z1 , z2 , . . . ) be any fixed
enumeration of all pairs of rational numbers. Let zi = (q1 , q2 ). Let u be the point in the plane with
coordinates (q1 , q2 ) in the system Σp . The port at p corresponding to edge {p, u} has number i.
Algorithm GeometricRV
The algorithm is a direct application of Algorithm GraphRV to the graph GT . The agent operating
in an unknown terrain T designs a route in the corresponding unknown graph GT as follows. When
the agent chooses a port at a node p ∈ V1 , this edge corresponds to some rational point qi in the
plane that the agent tries to reach from p. Two cases may occur. The agent either walks in the
interior of T until reaching qi , which corresponds to the traversal of an edge between two nodes of
V1 , or it hits the boundary of T , which corresponds to a visit of a node p0 ∈ V2 . At a node p0 ∈ V2
there is no choice of port, since its degree is 1. The agent takes the unique port which leads to
the (already visited) node p ∈ V1 . The resulting route is a sequence of segments joining rational
interior points of the terrain T and of pairs of consecutive segments (vb, bv), where v is a rational
interior point of T and b is a point on the boundary of T .
Since by Lemma 4.1 graph GT is connected, rendezvous is guaranteed in the graph GT , which
implies rendezvous in T .
8
Theorem 4.1 Algorithm GeometricRV guarantees asynchronous rendezvous for arbitrary two agents
starting from arbitrary rational interior points of any closed terrain T with path-connected interior.
Theorem 4.1 should be contrasted with the following negative result showing that the restriction
on the starting points of the agents cannot be removed, even for rendezvous in the (empty) plane.
Proposition 4.1 There is no algorithm that guarantees asynchronous rendezvous of arbitrary
agents starting from arbitrary points in the plane.
Proof: Consider the agent with label ` operating in the empty plane. Since the terrain is fixed,
the route of this agent depends only on the starting point. Let R = (e1 , e2 , ...) be the route of the
agent with label 1 starting at a fixed point v. Consider the route R2 (w) of the agent with label 2
starting at point w. Since there are no boundary points in the terrain, for any starting points w0
and w00 , route R2 (w00 ) is a parallel shift of route R2 (w0 ) by the vector (w0 , w00 ). Both of them are
polygonal lines.
We will say that two routes are almost disjoint if all vertices of each of them are outside the other
route. Observe that if, for some starting point w, route R2 (w) of agent 2 is almost disjoint from
route R of agent 1, then the adversary can avoid rendezvous of agents 1 and 2 following these
routes, by first moving agent 1 to the end of the first segment of its route, then moving agent 2 to
the end of the first segment of its route and so on, alternating traversals of agents on consecutive
segments of their routes. Hence, in order to prove our result, it is enough to show the existence of
a point w∗ , such that route R2 (w∗ ) is almost disjoint from route R.
Let (f1 , f2 , ...) be the sequence of vectors corresponding to consecutive segments of the route R2 (w),
for any starting point w. For any fixed point w, denote by fj (w) the segment corresponding to
vector fj on route R2 (w). Let (p1 , p2 , . . . ) be any sequence that orders all couples (ei , fj ), for all
positive integers i, j. For any k, let pk = (eik , fjk ) We will construct by induction a descending
sequence of closed discs (D1 , D2 , . . . ) of positive radii, satisfying the following invariant. For all
points w ∈ Dk , both endpoints of the segment eik are outside of the segment fjk (w) and both
endpoints of the segment fjk (w) are outside of the segment eik . Suppose that the invariant is
satisfied for k − 1 (for k = 1 we may take as D0 any disc with radius 1). The set of points w inside
disc Dk−1 that may possibly violate the invariant for k is contained in the union of four segments:
two segments parallel to fjk and two segments parallel to eik . There exists a closed disc of positive
radius contained in Dk−1 which is disjoint from those four segments. Let Dk be such a disc. Thus
the invariant is satisfied for k. This completes the construction by induction.
The intersection of all discs Dk is non-empty. Let w∗ be a point in this intersection. Since
(p1 , p2 , . . . ) enumerated all couples (ei , fj ), it follows that route R2 (w∗ ) is almost disjoint from
route R, and hence agents 1 and 2 starting at points v and w∗ , respectively, do not meet for some
walks chosen by the adversary.
While Proposition 4.1 shows that rendezvous of agents starting from arbitrary points is impossible,
it turns out that a slightly easier task can be accomplished in this setting. For any > 0, we say
that routes R1 and R2 of agents guarantee -approximate rendezvous, if at some point t of time, the
agents get at distance at most from each other, regardless of the walks chosen by the adversary.
Theorem 4.2 Algorithm GeometricRV guarantees -approximate rendezvous for any > 0, for
arbitrary agents starting from arbitrary interior points of any closed terrain T with path-connected
interior.
9
Proof: Fix an > 0. Consider any two agents starting at interior points v and w of the terrain
T . Let ρ > 0 be the distance from w to the boundary of T . Choose a point w0 with rational
coordinates in Σv , in the interior of the disc D of radius ρ/2 centered at w. By Lemma 4.1 (applied
to the system Σv instead of Σ), there exists a rational polygonal line P in the system Σv , included
in the interior of T and joining v and w0 . Let d > 0 be the distance between P and the boundary
of T . Let r = min(ρ/2, d/2, ). Choose a point w00 with rational coordinates in the system Σv , at
distance less than r from w. Let Pv be the polygonal line P extended by the segment w0 w00 . Note
that Pv is at distance at least r from the boundary of T . Indeed, if x ∈ P , then the distance from
x to the boundary is at least d > r, and if x ∈ w0 w00 , then the distance from x to the boundary is
at least ρ/2 ≥ r, as the entire segment w0 w00 is included in disc D.
Let Φ be the translation by the vector (w00 w) and let Pw be the image of Pv with respect to Φ. The
point w is one of the extremities of Pw . Since the distance from w00 to w is less than r, the entire
polygonal line Pw is in the interior of T . The polygonal line Pw is rational in the system Σw .
Let f be any walk of the first agent on any route with the initial part Pv starting at v, and let g
be any walk of the second agent on any route with the initial part Pw starting at w. Consider the
composition g ∗ = Φ−1 ◦ g. Hence g ∗ is a walk on a route with initial part Pv , starting at w00 . By
Theorem 4.1, Algorithm GeometricRV guarantees rendezvous of agents at some point in time t, for
the walk f starting at v and the walk g ∗ starting at w00 .
Consider the positions at time t of both agents starting at v and w, in walks f and g. Since
f (t) = g ∗ (t) = Φ−1 (g(t)), and Φ is a translation by a vector of length less than r (hence less than
), it follows that the distance between f (t) and g(t) is less than . Since walks f and g were
arbitrarily chosen by the adversary, this guarantees -approximate rendezvous.
A consequence of Theorem 4.2 is that if agents have arbitrarily small positive visibility ranges
(rather than 0 visibility range as we assumed) and they start in arbitrary points of the (empty)
plane, then Algorithm GeometricRV guarantees that they will see each other in finite time, regardless of the actions of the adversary.
5
Conclusion
We provided deterministic asynchronous rendezvous algorithms for graphs and for terrains in the
plane. We studied only the feasibility of rendezvous and our results are very general: for the graph
scenario, we showed that rendezvous is possible in any connected countable (finite or infinite) graph,
starting from any nodes, without any information on the graph. The only thing an agent needs to
know is its own label. In particular, this result implies a positive solution of a problem from [12].
Our algorithms rely on an arbitrary fixed enumeration of quadruples (i, j, s0 , s00 ), where i and
j are positive integers and s0 and s00 are finite sequences of positive integers. The complexity
of the algorithm (measured by the worst-case length of paths that the agents have to traverse
until rendezvous) depends on this enumeration, but we do not think any enumeration can make
the algorithm efficient. Thus a natural interesting question left for further investigations is the
following:
Does there exist a deterministic asynchronous rendezvous algorithm, working for all
connected countable unknown graphs, with complexity polynomial in the labels of the
agents and in the initial distance between them?
10
References
[1] R. Aleliunas, R.M. Karp, R.J. Lipton, L. Lov´asz, and C. Rackoff, Random walks, universal traversal sequences, and the complexity of maze problems, Proc. Annual Symposium on
Foundations of Computer Science FOCS’1979, 218-223.
[2] S. Alpern, The rendezvous search problem, SIAM J. on Control and Optimization 33 (1995),
673-683.
[3] S. Alpern, Rendezvous search on labelled networks, Naval Reaserch Logistics 49 (2002), 256274.
[4] S. Alpern and S. Gal, The theory of search games and rendezvous. Int. Series in Operations
research and Management Science, number 55, Kluwer Academic Publishers, 2002. Kluwer
Academic Publisher, 2002.
[5] J. Alpern, V. Baston, and S. Essegaier, Rendezvous search on a graph, Journal of Applied
Probability 36 (1999), 223-231.
[6] E. Anderson and R. Weber, The rendezvous problem on discrete locations, Journal of Applied
Probability 28 (1990), 839-851.
[7] E. Anderson and S. Fekete, Asymmetric rendezvous on the plane, Proc. 14th Annual ACM
Symp. on Computational Geometry, 365-373, 1998.
[8] E. Anderson and S. Fekete, Two-dimensional rendezvous search, Operations Research 49
(2001), 107-118.
[9] V. Baston and S. Gal, Rendezvous on the line when the players’ initial distance is given by an
unknown probability distribution, SIAM J. on Control and Optimization 36 (1998), 1880-1889.
[10] V. Baston and S. Gal, Rendezvous search when marks are left at the starting points, Naval
Res. Log. 48 (2001), 722-731.
[11] M. Cieliebak, P. Flocchini, G. Prencipe, N. Santoro, Solving the Robots Gathering Problem,
Proc. 30th International Colloquium on Automata, Languages and Programming (ICALP
2003), 1181-1196.
[12] G. De Marco, L. Gargano, E. Kranakis, D. Krizanc, A. Pelc, U. Vaccaro, Asynchronous deterministic rendezvous in graphs, Theoretical Computer Science 355 (2006), 315-326.
[13] D. Coppersmith,, P. Doyle, P. Raghavan, and M. Snir, Random walks on weighted graphs,
and applications to on-line algorithms, Proc. 22nd Annual ACM Symposium on Theory of
Computing ( STOC’1990), 369-378.
[14] A. Dessmark, P. Fraigniaud, D. Kowalski, and A. Pelc, Deterministic rendezvous in graphs,
Algorithmica 46 (2006), 69-96.
[15] M. Dyer, A. Frieze, and R. Kannan, A random polynomial time algorithm for estimating
volumes of convex bodies, Proc. 21st Annual ACM Symposium on Theory of Computing
(STOC’1989), 375-381.
11
[16] P. Flocchini, G. Prencipe, N. Santoro, P. Widmayer, Gathering of asynchronous oblivious
robots with limited visibility, Proc. 18th Annual Symposium on Theoretical Aspects of Computer Science STACS’2001, LNCS 2010, 247-258.
[17] S. Gal, Rendezvous search on the line, Operations Research 47 (1999), 974-976.
[18] A. Israeli and M. Jalfon, Token management schemes and random walks yield self stabilizing
mutual exclusion, Proc. PODC’1990, 119-131.
[19] R. Klasing, A. Kosowski, A. Navarra, Taking advantage of symmetries: gathering of asynchronous oblivious robots on a ring. Proc. 12th International Conference on Principles of
Distributed Systems, (OPODIS 2008), 446-462.
[20] R. Klasing, E. Markou, A. Pelc, Gathering asynchronous oblivious mobile robots in a ring,
Theoretical Computer Science 390 (2008), 27-39.
[21] D. Kowalski, A. Malinowski, How to meet in anonymous network. Theoretical Computer Science 399 (2008), 141-156.
[22] E. Kranakis, D. Krizanc, N. Santoro and C. Sawchuk, Mobile agent rendezvous in a ring, Proc.
23rd International Conference on Distributed Computing Systems (ICDCS’2003), 592-599.
[23] W. Lim and S. Alpern, Minimax rendezvous on the line, SIAM J. on Control and Optimization
34 (1996), 1650-1665.
[24] G. Prencipe, Impossibility of gathering by a set of autonomous mobile robots, Theoretical
Computer Science 384 (2007), 222-231.
[25] T. Schelling, The strategy of conflict, Oxford University Press, Oxford, 1960.
[26] G. Stachowiak, Asynchronous Deterministic Rendezvous on the Line, SOFSEM 2009: 497-508.
[27] L. Thomas, Finding your kids when they are lost, Journal on Operational Res. Soc. 43 (1992),
637-639.
[28] X. Yu and M. Yung, Agent rendezvous: a dynamic symmetry-breaking problem, Proc. International Colloquium on Automata, Languages, and Programming (ICALP’1996), LNCS 1099,
610-621.
[29] A. Ta-Shma, U. Zwick, Deterministic rendezvous, treasure hunts and strongly universal exploration sequences., Proc. 18th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA
2007), 599-608.
12